SahulatAI

Autonomous AI agents powered by Gemini that run small business operations end-to-end using WhatsApp, spreadsheets, and real-world tools.

Comment

Inspiration

Small and informal merchants in Pakistan run their businesses on WhatsApp and spreadsheets. During the National Agentic AI Week and dozens of merchant interviews, we watched owners lose customers because replies were slow, inventory was inconsistent, and marketing was manual. Existing business software is either too costly or too complex. We built SahulatAI to give these businesses an AI workforce — not another dashboard — so work gets done for them automatically. Gemini’s reasoning and multimodal capabilities made this vision practical.

What it does

SahulatAI is a multi-tenant, and multi-directional agentic platform that automates core business operations for small merchants. Each autonomous agent has a role (Sales, Support, Inventory, Marketing, Payments, Accounts) and can:

  • Understand WhatsApp conversations and turn them into orders or support actions.
  • Update inventory in Google Sheets and trigger restock or low-stock alerts.
  • Generate marketing posters and short ad copy using Gemini’s multimodal reasoning.
  • Create payment links and follow up on orders.
  • Perform bookkeeping actions (invoicing, recording payments) via secure, tenant-isolated MCP tooling.

In short: merchants onboard in minutes (QR + Sheets) and the agents run their day-to-day operations.

How we built it

Architecture & orchestration

  • Agent orchestration layer (FastAPI) runs coordinated workflows and lifecycle tracing.
  • Tenant-isolated MCP servers handle secure OAuth and tool invocations (QuickBooks, Sheets, payment APIs).
  • PostgreSQL stores conversation context, agent state, and logs.
  • Frontend: React / Next.js for onboarding & admin.
  • Deployment: GCP & others

Gemini role (core)

  • Conversation understanding & planning: Gemini 3 Pro for WhatsApp messages, extracts intents, and plans multi-step workflows.
  • Multimodal content generation: Gemini’s vision+reasoning powers poster generation and image captions.
  • Tool selection & execution: Gemini suggests which tool/agent to call and provides the reasoning trace for safe execution.

Integrations

  • WhatsApp Opensource API for conversational I/O
  • Google Sheets API for inventory sync
  • QuickBooks API for accounting actions
  • Image CDN for poster storage

Challenges we ran into

  • Messy real-world data: User sheets vary wildly in format; we built a data-sanitization layer and heuristics.
  • Reliable action execution: Agents performing real actions require tracing, retry logic, and strict guardrails to avoid bad writes.
  • Cost & latency: Multi-agent workflows can trigger many model calls; we optimized by caching, batching, and using smaller models for routine tasks.
  • Secure integrations: QuickBooks and payment flows required tenant isolation and secure token handling via MCP servers.

Accomplishments that we're proud of

  • Built a functioning multi-agent MVP that executes real workflows end-to-end.
  • Demonstrated live WhatsApp → order → inventory update → invoice flow in demos with local merchants.
  • Validated customer interest with multiple demo sessions and pilot requests following hackathon exposure.
  • Successfully integrated multimodal poster generation powered by Gemini to help merchants create marketing content in seconds.

What we learned

  • Merchants want outcomes, not dashboards — automation that does the work is the key to adoption.
  • Onboarding must be as simple as possible (QR + Sheets) to reduce friction.
  • Observability (traces, logs, human-review queues) is essential when agents act on behalf of users.
  • Choosing the right mix of large and small models dramatically reduces cost while preserving capability.

What's next for SahulatAI

  • Expand Gemini-driven capabilities: more robust multimodal ads, improved long-context planning.
  • Roll out multi-channel support (Instagram DMs, Facebook Messenger).
  • Add supplier discovery (nearest & best supplier suggestions) and marketplace workflows.
  • Run structured pilots (50–200 merchants) to validate pricing tiers and reduce CAC.
  • Optimize model usage and build an analytics dashboard for ROI tracking.

Built With

Gemini 3 Pro, Gemini Nano Banana
FastAPI, React / Next.js
Google Cloud Platform (Compute, Cloud Run)
PostgreSQL
WhatsApp Cloud API
Google Sheets API
QuickBooks API
Image CDN (for poster storage)
Datadog (observability)

Try it out

Built With

Share this project:

Updates